Appearance-Based Driver 3-D Gaze Estimation Using GRM and Mixed Loss Strategies

Driver gaze estimation is a key technology in advanced intelligent vehicles, and it is crucial for ensuring road safety by monitoring driver visual attention. Previously, attention detection through head pose or saliency map integration only offered rudimentary estimation and was insufficient for the advanced driver assistance systems (ADASs), which require more precise gaze data. This work introduces an appearance-based method for driver 3-D gaze estimation. Initially, the Swin Transformer was used to enhance global image information processing, which enabled accurate gaze direction prediction. Furthermore, the method incorporates a gaze refinement module (GRM) as a postbackbone to optimize feature mapping, thus ensuring stable gaze direction estimation. Finally, a mixed loss function was used to improve the accuracy. This mixed loss function combines pinball loss, mean-squared error (MSE), and bias penalty. The experimental results demonstrated angular errors of 3.76° and 10.62° in the MPIIGazeFace and Gaze360 gaze estimation data sets. We inferenced the proposed method to the driver monitoring data set (DMD), and the results demonstrate the effectiveness of this work. Our code is publicly available at github.com/Rocky1salady-killer/DGE-GM.